US20230376825A1 - Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift - Google Patents
Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift Download PDFInfo
- Publication number
- US20230376825A1 US20230376825A1 US17/663,848 US202217663848A US2023376825A1 US 20230376825 A1 US20230376825 A1 US 20230376825A1 US 202217663848 A US202217663848 A US 202217663848A US 2023376825 A1 US2023376825 A1 US 2023376825A1
- Authority
- US
- United States
- Prior art keywords
- drift
- model
- data
- score
- concept
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 97
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 35
- 238000000034 method Methods 0.000 claims abstract description 60
- 230000006870 function Effects 0.000 claims abstract description 53
- 230000004044 response Effects 0.000 claims abstract description 52
- 238000004590 computer program Methods 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims description 48
- 238000003860 storage Methods 0.000 claims description 32
- 238000010586 diagram Methods 0.000 description 20
- 238000012545 processing Methods 0.000 description 12
- 238000004458 analytical method Methods 0.000 description 8
- 238000004519 manufacturing process Methods 0.000 description 7
- 238000007726 management method Methods 0.000 description 5
- 230000007246 mechanism Effects 0.000 description 5
- 238000005457 optimization Methods 0.000 description 5
- 238000010200 validation analysis Methods 0.000 description 5
- 230000005540 biological transmission Effects 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 230000008520 organization Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000003491 array Methods 0.000 description 2
- 230000015556 catabolic process Effects 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000006855 networking Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000001902 propagating effect Effects 0.000 description 2
- 238000012384 transportation and delivery Methods 0.000 description 2
- 241000283690 Bos taurus Species 0.000 description 1
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- 230000003466 anti-cipated effect Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009172 bursting Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000012517 data analytics Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000013439 planning Methods 0.000 description 1
- 229920001690 polydopamine Polymers 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000013468 resource allocation Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Definitions
- the present invention relates generally to artificial intelligence models, and more particularly to adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift.
- AI enterprise artificial intelligence
- IoT industrial Internet of Things
- multiple AI models are prepared and deployed to fulfill business requirements. Monitoring the behavior of these AI models and retrain as the model performance degrades by identifying suitable retraining data is essential to the business. Analysis of the models alone is not sufficient to decide on whether to retrain or to select data for the retraining. It is also required to identify changes in input data (features) and targets (ground truth) to perform retraining.
- various algorithms exist to identify various types of drifts, e.g., a data drift, a concept (target) drift, and a model drift, a system that enables an automated analysis of these drifts and the relationship between them is required for business personnel.
- a computer-implemented method for adaptive retraining of an artificial intelligence model includes computing drift magnitude scores for respective drift functions.
- the computer-implemented method further includes computing an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift.
- the computer-implemented method includes computing an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold.
- the computer-implemented method includes determining whether retraining of the artificial intelligence model is required, based on the overall drift score.
- the computer-implemented method includes performing the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
- a computer program product for adaptive retraining of an artificial intelligence model.
- the computer program product comprises a computer readable storage medium having program instructions embodied therewith, and the program instructions are executable by one or more processors.
- the program instructions are executable to: compute drift magnitude scores for respective drift functions; compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift; compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold; determine whether retraining of the artificial intelligence model is required, based on the overall drift score; and perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
- a computer system for adaptive retraining of an artificial intelligence model comprises one or more processors, one or more computer readable tangible storage devices, and program instructions stored on at least one of the one or more computer readable tangible storage devices for execution by at least one of the one or more processors.
- the program instructions are executable to compute drift magnitude scores for respective drift functions.
- the program instructions are further executable to compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift.
- the program instructions are further executable to compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold.
- the program instructions are further executable to determine whether retraining of the artificial intelligence model is required, based on the overall drift score.
- the program instructions are further executable to perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
- FIG. 1 is a diagram illustrating a problem setting of a predictive model, in accordance with one embodiment of the present invention.
- FIG. 2 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with one embodiment of the present invention.
- FIG. 3 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- FIG. 4 illustrates an example of relationships between a data drift score, a concept drift score, a model drift score, and an overall drift score, in accordance with another embodiment of the present invention.
- FIG. 5 is a diagram illustrating a first example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- FIG. 6 is a diagram illustrating a second example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- FIG. 7 is a diagram illustrating a third example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- FIG. 8 is a diagram illustrating components of a computing device or server, in accordance with one embodiment of the present invention.
- FIG. 9 depicts a cloud computing environment, in accordance with one embodiment of the present invention.
- FIG. 10 depicts abstraction model layers in a cloud computing environment, in accordance with one embodiment of the present invention.
- FIG. 1 is a diagram illustrating a problem setting of predictive model 110 , in accordance with one embodiment of the present invention.
- Input sensor data (X) 120 is generated in periodic interval from multiple sensors installed in various assets and operations.
- Feature extracted data (X′) 130 is prepared from timeseries of input sensor data (X) 120 using feature extraction techniques.
- Predictive model or AI model (M) 110 may use either input sensor data (X) 120 or feature extracted data (X′) 130 for training.
- Predictive model or AI model (M) 110 is trained on input sensor data (X) 120 or feature extracted data (X′) 130 and outputs predictions ( ⁇ ) 140 in example 1 and/or predictions ( ⁇ ) 150 in example 2 for all the subsequent input data.
- predictions ( ⁇ ) 140 is classification output; in example 2, predictions ( ⁇ ) 150 is regression output.
- ground truth (Y) 160 is available for predictive model (M) 110 ; in example 2, ground truth (Y) 170 is available for predictive model (M) 110 .
- the AI system has following drifts.
- the data drift is a drift in input data.
- the analysis of the data drift requires only the input data, e.g., timeseries of input sensor data (X) 120 or feature extracted data (X′) 130 .
- the concept drift is a drift in target labels (e.g., ground truth (Y) 160 or ground truth (Y) 170 ) or the relationship between target labels and input data.
- the model drift is a drift in predictions of predictive model or AI model (M) 110 , the relationship between the predictions (e.g., predictions ( ⁇ ) 140 and/or predictions ( ⁇ ) 150 ) and the target labels, or the relationship between the input data, the predictions, and the target labels.
- the analysis of the model drift may include comparison of model parameters or optimization parameters emitted during model preparation.
- a client model (such as predictive model (M) 110 ) is required in order to analyze the model drift.
- Drawbacks of the concept drift driven adaptive retraining are as follows: (1) The concept drift alone is not the reason for model degradation. (2) There is no mechanism to accommodate multiple drift functions from other drift types. In the quality driven adaptive retraining, model quality is observed over a time period and the abrupt model drift is identified using moving average techniques. In identifying the model quality, model predictions against ground truth are analyzed. The quality driven adaptive retraining has following drawbacks: (1) The reason for the model drift is not explained. (2) There is no mechanism to accommodate multiple drift functions. (3) There is no mechanism or method to anticipate possible model degradation prior to happening. (4) The drift in ground truth (or the concept drift) is not addressed as well.
- Embodiments of the present invention disclose a system and method for quantifying different types of drifts in training artificial intelligence (AI) models, e.g., the data drift, the concept drift, and the model drift, using proper drift functions/algorithms.
- AI artificial intelligence
- the disclosed system and method detect the drifts and notify users to retrain the AI models when required, using combined analysis of the above drifts.
- the disclosed system and method select appropriate data required for retraining using the above analysis and perform retraining of the AI model.
- the disclosed system and method train a model on given training data, deploy the model in production, monitor the model over the time, and retrain the model as required.
- the disclosed system comprises two major modules: a model training module and an adaptive retraining module.
- the disclosed system and method are implemented on one or more computing devices or servers.
- a computing device or server is described in more detail in later paragraphs with reference to FIG. 8 .
- the disclosed system and method may be implemented in a cloud computing environment.
- the cloud computing environment is described in more detail in later paragraphs with reference to FIG. 9 and FIG. 10 .
- the model training module trains and deploys a predictive model or AI model. From a pool of pre-selected models, the model training module chooses a best model suitable for a given dataset and a problem, using manual analysis or automated selection such as optimization-based selection and grid search. The model training module determines model parameters for the best model, using hyperparameter search. The model training module trains the best model using training data and ground truth. Part of the training data may be used as validation data. The model training module deploys the best model in production. Test data is scored using the deployed best model and predictions are obtained by the model training module. The deployed best model is selected to be monitored by the adaptive retrain module. The implementation of the adaptive retrain module includes two phases: phase 1—computing retraining score, and phase 2—identifying required retraining data and creating a new AI model.
- Embodiments of the present invention disclose an automated system that perform adaptive retraining by analyzing the data drift, the concept drift, and the model drift together.
- an overall drifting score or a retraining score which considers the relationship between the data drift, the concept drift, and the model drift, is computed.
- Embodiments of the present invention develop a 3-level drift scoring technique. First, a score for each drift function that resembles magnitude of change occurred in the data is computed. Second, a score for each of the data drift, the concept drift, and the model drift by aggregation of scores from the respective drift functions.
- the overall drift score (or the retraining score) is computed by considering the second-level scores, i.e., an aggregated data drift score, an aggregated concept drift score, and an aggregated model drift score.
- the second-level scores i.e., an aggregated data drift score, an aggregated concept drift score, and an aggregated model drift score.
- Embodiments of the present invention select appropriate retraining data by a combination of sampling and relabeling techniques, based on the 3-level drift scores.
- Embodiments of the present invention allow users to take an informed decision in automated manner using combined analysis of data, concept, and model drifts.
- any algorithm which is used to identify a drift in input data, a drift in ground truth, a drift in model predictions, or relationship among them is called a drift function.
- the drift function is categorized into one of the categories: the data drift functions, the concept drift functions, and the model drift functions.
- the proposed system and method re-categorize all the available drift functions ⁇ 1 , ⁇ 2 , ⁇ 3 , . . . ⁇ n into data drift functions, concept drift functions, and model drift functions as per drift definitions.
- Drift thresholds including a drift threshold ( ⁇ ⁇ i ) for a drift function ⁇ i , a data drift threshold ( ⁇ dd ), a concept drift threshold ( ⁇ cd ), and a model drift threshold ( ⁇ md ), can be set either by predefined values or by users.
- the inputs include a batch of test data, model predictions, and ground truth.
- the outputs in the second-level drift score computation include an aggregated data drift score (s dd ) and its drift flag ( ⁇ dd ), an aggregated concept drift score (s cd ) and its drift flag ( ⁇ cd ), and an aggregated model drift score ( ⁇ md ) and its drift flag ( ⁇ md ).
- the outputs in the third-level drift score computation include a retraining score (s) and a retraining flag ⁇ .
- the retraining score (s) may also called as an overall drift score or a total drift score.
- the data elements from the model training phase i.e., training data, validation data, model predictions for training or validation data, ground truth, model parameters, optimization statistics, etc., can be used as a base in the adaptive retrain module. Computation of the 3-level drift scores will be discussed in detail in later paragraphs of this document.
- the retraining score (s) and retraining flag ⁇ are obtained. Based on the retraining score (s) and retraining flag ⁇ , the disclosed system and method determine whether retraining of predictive model or AI model is required. If the retraining flag ⁇ is true, the disclosed system and method perform the retraining.
- the aggregated data drift score (s dd ), the data drift flag ( ⁇ dd ), the aggregated concept drift score (s cd ), the concept drift flag ( ⁇ cd ), the aggregated model drift score (s md ), and the model drift flag ( ⁇ cd ) are obtained.
- the disclosed system and method determines that the data drift is present.
- the disclosed system and method determines that the concept drift is present.
- the disclosed system and method determines that the model drift flag ( ⁇ cd ) is true.
- phase 2 identifying required retraining data and creating a new AI model
- the disclosed system and method select new training data from the drifted period using known sampling techniques.
- the disclosed system and method use the original training data. Different strategies can be used to prepare new training data from the drifted period.
- phase 2 identifying required retraining data and creating a new AI model
- the disclosed system and method apply known relabeling techniques on the selected training data.
- the disclosed system and method may take new label definitions as inputs or deduce new label definitions from the drifted data.
- phase 2 identifying required retraining data and creating a new AI model
- model selection is ignored by the disclosed system and method. Otherwise, the disclosed system and method perform model selection to select a new AI model. Finally, hyperparameter search is performed on the selected new AI model.
- the disclosed system and method may replace the previously deployed AI model with the new AI model.
- the disclosed system and method may compare both the previously deployed AI model with the new AI model against each other using drift results over future period of data, and then the disclosed system and method select a winner as the production model.
- a drift function ⁇ i has a threshold ⁇ ⁇ i (>0) set to either a user provided value or a default value.
- ⁇ i computes metrics for a given batch of data either by comparing against training data or by comparing against recent data over a window or any other means.
- Example of a metric is standard deviation (std).
- a drift magnitude score (s ⁇ i ) and a drift flag ( ⁇ ⁇ i ) is computed for ⁇ i .
- the drift magnitude score (s ⁇ i ) and drift flag ( ⁇ ⁇ i ) are computed by
- drift magnitude score (s ⁇ i ) and drift flag ( ⁇ ⁇ i ) are computed for all data drift functions, concept drift functions, and model drift functions ⁇ 1 , ⁇ 2 , ⁇ 3 , . . . ⁇ n .
- Each of the data drift, the concept drift, and the model drift has at least one drift function.
- the data drift, the concept drift, and the model drift have thresholds ⁇ dd , ⁇ cd , and ⁇ md , respectively; the thresholds can be set either by predefined values or by users.
- the data drift has drift function ⁇ 1 , ⁇ 2 , . . . ⁇ m which evaluate either univariate or multivariate input data.
- the aggregated data drift score (s dd ) and the data drift flag ( ⁇ dd ) are computed from drift magnitude scores and drift flags corresponding to all the data drift functions ⁇ 1 , ⁇ 2 , . . . ⁇ m .
- the aggregated data drift score (s dd ) and the data drift flag ( ⁇ dd ) are computed by
- the aggregated concept drift score (s cd ), the concept drift flag ( ⁇ cd ), the aggregated model drift score (s md ), and model drift flag ( ⁇ md ) are computed in a similar way as the aggregated data drift score (s dd ) and the data drift flag ( ⁇ dd ).
- the retraining score or overall drift score (s) and the retraining flag ⁇ are computed based on the aggregated data drift score (s dd ) and its drift flag ( ⁇ dd ), the aggregated concept drift score (s cd ) and its drift flag ( ⁇ cd ), and the aggregated model drift score (s md ) and its drift flag ( ⁇ md ).
- the aggregated data drift score (s dd ), the aggregated concept drift score (s cd ), and the aggregated model drift score (s md ) can be aggregated values over a period.
- An example of aggregation is the average of the values.
- the retraining score or overall drift score (s) and the retraining flag ⁇ are computed by
- r dd s dd ⁇ dd ( 5 )
- r c ⁇ d s c ⁇ d ⁇ c ⁇ d ( 6 )
- r md s md ⁇ md ( 7 )
- s ⁇ r md ⁇ r c ⁇ d , if ⁇ r dd ⁇ 1 ⁇ and ⁇ r c ⁇ d ⁇ 1.
- r md ⁇ r dd if ⁇ r c ⁇ d ⁇ 1 ⁇ and ⁇ r md ⁇ 1.
- FIG. 2 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with one embodiment of the present invention.
- the operational steps presented in FIG. 2 are implemented by the disclosed system which is hosted by one or more computing devices or servers.
- a computing device or server is described in more detail in later paragraphs with reference to FIG. 8 .
- the operational steps presented in FIG. 2 may be implemented in a cloud computing environment.
- the cloud computing environment is described in more detail in later paragraphs with reference to FIG. 9 and FIG. 10 .
- the one or more computing devices or servers select an artificial intelligence (AI) model for a dataset and a problem. From a pool of pre-selected models, the one or more computing devices or servers choose a best model suitable for the given dataset and the problem.
- the one or more computing devices or servers determine parameters for the AI model using hyperparameter search.
- the one or more computing devices or servers train the AI model using training data and ground truth.
- the one or more computing devices or servers deploy the AI model and obtain predictions using the AI model.
- Steps 201 - 204 are operational steps in the model training phase.
- the data elements from the model training phase including training data, validation data, model predictions for training or validation data, ground truth, model parameters, and optimization statistics, can be used to compute 3-level drift scores in the following steps.
- the one or more computing devices or servers compute drift magnitude scores for respective drift functions.
- Drift functions ⁇ 1 , ⁇ 2 , ⁇ 3 , . . . ⁇ n are available, each of which is an algorithm used to identify a drift in input data, a drift in ground truth, a drift in model predictions, or a relationship among them.
- the one or more computing devices or servers categorize all the available drift functions ⁇ 1 , ⁇ 2 , ⁇ 3 , . . . ⁇ n into data drift functions, concept drift functions, and model drift functions.
- a drift magnitude score (s ⁇ i ) the one or more computing devices or servers compute a drift magnitude score (s ⁇ i ), using equation 1 described in a previous paragraph in this document.
- the one or more computing devices or servers also compute a drift flag ( ⁇ ⁇ i ), using equation 2 described in a previous paragraph in this document.
- the one or more computing devices or servers compute the first-level drift scores.
- the one or more computing devices or servers compute an aggregated data drift score, an aggregated concept drift score, and an aggregated model drift score. Based on the drift magnitude score (s ⁇ i ) and the drift flag ( ⁇ ⁇ i ) (the first-level drift scores) computed at step 205 , the one or more computing devices or servers compute the aggregated data drift score (s dd ), using equation 3 described in a previous paragraph in this document. The one or more computing devices or servers at step 206 also compute the data drift flag ( ⁇ dd ), using equation 4 described in a previous paragraph in this document, based on the aggregated data drift score (s dd ) and a predetermined threshold ( ⁇ dd ) of the data shift.
- the one or more computing devices or servers compute the aggregated concept drift score (s cd ), the concept drift flag ( ⁇ cd ), the aggregated model drift score (s md ), and model drift flag ( ⁇ cd ).
- the one or more computing devices or servers compute the second-level drift scores.
- the one or more computing devices or servers compute a retraining score (or an overall drift score), based on the aggregated data drift score, the aggregated concept drift score, and the aggregated model drift score. Based on the aggregated data drift score (s dd ), the aggregated concept drift score (s cd ), and aggregated model drift score (s md ) that are computed at step 206 , and also based on the predetermined threshold ( ⁇ dd ) of the data shift, the predetermined threshold ( ⁇ cd ) of the concept shift, and the predetermined threshold (rind) of the model drift, the one or more computing devices or servers compute the retraining score or overall drift score (s), using equations 5-9 described in a previous paragraph in this document. At step 207 , the one or more computing devices or servers also compute a retraining flag ( ⁇ ), using equation 10 described in a previous paragraph in this document. At step 207 , the one or more computing devices or servers compute the third-level drift scores.
- the one or more computing devices or servers determine whether retaining of the AI model is required, based on the retraining score (the retraining score may also called as an overall drift score or a total drift score). To determine whether the retaining of the AI model is required, the one or more computing devices or servers determines whether the retraining flag ⁇ is true or false.
- the retraining score or overall drift score (s) and retraining flag ⁇ have been computed at step 207 by the one or more computing devices or servers.
- the one or more computing devices or servers In response to determining that the retaining the AI model is not required (NO branch of decision block 209 ), the one or more computing devices or servers does not perform the retraining of the AI model. In other words, in response to determining that the retraining flag ⁇ is false, the one or more computing devices or servers will not take any action to retrain the AI model.
- the one or more computing devices or servers perform the retraining of the AI model. In other words, in response to determining that the retraining flag ⁇ is true, the one or more computing devices or servers perform the retraining.
- FIG. 3 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- the operational steps presented in FIG. 3 are implemented by the disclosed system which is hosted by one or more computing devices or servers.
- a computing device or server is described in more detail in later paragraphs with reference to FIG. 8 .
- the operational steps may be implemented in a cloud computing environment.
- the cloud computing environment is described in more detail in later paragraphs with reference to FIG. 9 and FIG. 10 .
- the aggregated data drift score, aggregated concept drift score, and aggregated model drift score have been computed at step 206 by the one or more computing devices or servers.
- the aggregated data drift score, aggregated concept drift score, and aggregated model drift score are used for the operating steps presented in FIG. 3 .
- the one or more computing devices or servers determine whether the data drift is present. To determine whether the data drift is present, the one or more computing devices or servers determine whether the data drift flag ( ⁇ dd ) (which is computed at step 206 ) is true or false. If the data drift flag ( ⁇ dd ) is true, the data drift is present; if the data drift flag ( ⁇ dd ) is false, the data drift is not present.
- the data drift flag ( ⁇ dd ) which is computed at step 206 .
- the one or more computing devices or servers use new training data for the retraining of the AI model.
- the new training data may be an entirely new training dataset.
- the new training data may be combination of the drifted training data and non-drifted training data (or the one or more computing devices or servers incorporate the drifted training data into new data). Selecting the new training data from the drifted period uses known sampling techniques.
- the one or more computing devices or servers use original data which has been used in the model training phase.
- the one or more computing devices or servers at step 304 determine whether the concept drift is present. To determine whether the concept drift is present, the one or more computing devices or servers determine whether the concept drift flag ( ⁇ cd ) (which is computed at step 206 ) is true or false. If the concept drift flag ( ⁇ dd ) is true, the concept drift is present; if the concept drift flag ( ⁇ cd ) is false, the concept drift is not present.
- the one or more computing devices or servers incorporate drifted targets into a dataset for the retraining of the AI model.
- the one or more computing devices or servers determine whether the model drift is present. To determine whether the model drift is present, the one or more computing devices or servers determine whether the model drift flag ( ⁇ cd ) (which is computed at step 206 ) is true or false. If the model drift flag ( ⁇ cd ) is true, the model drift is present; if the model drift flag ( ⁇ cd ) is false, the model drift is not present.
- the one or more computing devices or servers determine whether either data drift or the concept drift is present. In response to determining that the concept drift is not present or the model drift flag ( ⁇ cd ) is false (NO branch of decision block 306 ), at step 308 , the one or more computing devices or servers determine whether the concept drift is present.
- the one or more computing devices or servers at step 310 perform hyperparameter search.
- the one or more computing devices or servers at step 309 perform model selection and then perform hyperparameter search.
- the one or more computing devices or servers at step 309 perform model selection and then perform hyperparameter search.
- the one or more computing devices or servers will not take any action to retrain the AI model.
- FIG. 4 illustrates an example of relationships between a data drift score, a concept drift score, a model drift score, and an overall drift score, in accordance with another embodiment of the present invention.
- a data drift, a concept drift, and a model drift are present (i.e., data drift flag ( ⁇ dd ), concept drift flag ( ⁇ cd ), and model flag ( ⁇ md ) are true)
- the overall drift score which is calculated based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score
- the disclosed system and method incorporate drifted input data and drifted targets into new training data, and the disclosed system and method perform hyperparameter search.
- the disclosed system and method incorporate drifted input data and drifted targets into new training data, the disclosed system and method perform new model selection.
- the disclosed system and method incorporate drifted input data into new training data and perform hyperparameter search.
- the disclosed system and method incorporate drifted targets into new training data and perform hyperparameter search.
- the disclosed system and method perform new model selection; for example, the disclosed system and method select a different model from pipeline optimization.
- FIG. 5 is a diagram illustrating a first example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- An AI model M 0 is trained on a historic training data, by using model selection and hyperparameter search.
- drift algorithms are applied on the model M 0 and the data.
- FIG. 6 is a diagram illustrating a second example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- drift algorithms are applied on the model M 0 and the data.
- FIG. 7 is a diagram illustrating a third example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention.
- drift algorithms are applied on the model M 0 and the data.
- FIG. 8 is a diagram illustrating components of computing device or server 800 , in accordance with one embodiment of the present invention, in accordance with one embodiment of the present invention. It should be appreciated that FIG. 8 provides only an illustration of one implementation and does not imply any limitations; different embodiments may be implemented.
- computing device or server 800 includes processor(s) 820 , memory 810 , and tangible storage device(s) 830 .
- communications among the above-mentioned components of computing device or server 800 are denoted by numeral 890 .
- Memory 810 includes ROM(s) (Read Only Memory) 811 , RAM(s) (Random Access Memory) 813 , and cache(s) 815 .
- One or more operating systems 831 and one or more computer programs 833 reside on one or more computer readable tangible storage device(s) 830 .
- Computing device or server 800 further includes I/O interface(s) 850 .
- I/O interface(s) 850 allows for input and output of data with external device(s) 860 that may be connected to computing device or server 800 .
- Computing device or server 800 further includes network interface(s) 840 for communications between computing device or server 800 and a computer network.
- the present invention may be a system, a method, and/or a computer program product at any possible technical detail level of integration
- the computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention
- the computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device.
- the computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
- a non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing.
- RAM random access memory
- ROM read-only memory
- EPROM or Flash memory erasable programmable read-only memory
- SRAM static random access memory
- CD-ROM compact disc read-only memory
- DVD digital versatile disk
- memory stick a floppy disk
- a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon
- a computer readable storage medium is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.
- Computer readable program instructions described herein can be downloaded to respective computing/processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and/or a wireless network.
- the network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers.
- a network adapter card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing/processing device.
- Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, and procedural programming languages, such as the C programming language or similar programming languages.
- the computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
- the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).
- electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.
- These computer readable program instructions may be provided to a processor of a computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
- These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function/act specified in the flowchart and/or block diagram block or blocks.
- the computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions/acts specified in the flowchart and/or block diagram block or blocks.
- each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s).
- the functions noted in the blocks may occur out of the order noted in the Figures.
- two blocks shown in succession may, in fact, be accomplished as one step, executed concurrently, substantially concurrently, in a partially or wholly temporally overlapping manner, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.
- Cloud computing is a model of service delivery for enabling convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines, and services) that can be rapidly provisioned and released with minimal management effort or interaction with a provider of the service.
- This cloud model may include at least five characteristics, at least three service models, and at least four deployment models.
- On-demand self-service a cloud consumer can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with the service's provider.
- Resource pooling the provider's computing resources are pooled to serve multiple consumers using a multi-tenant model, with different physical and virtual resources dynamically assigned and reassigned according to demand. There is a sense of location independence in that the consumer generally has no control or knowledge over the exact location of the provided resources but may be able to specify location at a higher level of abstraction (e.g., country, state, or datacenter).
- Rapid elasticity capabilities can be rapidly and elastically provisioned, in some cases automatically, to quickly scale out and rapidly released to quickly scale in. To the consumer, the capabilities available for provisioning often appear to be unlimited and can be purchased in any quantity at any time.
- Measured service cloud systems automatically control and optimize resource use by leveraging a metering capability at some level of abstraction appropriate to the type of service (e.g., storage, processing, bandwidth, and active user accounts). Resource usage can be monitored, controlled, and reported, providing transparency for both the provider and consumer of the utilized service.
- level of abstraction appropriate to the type of service (e.g., storage, processing, bandwidth, and active user accounts).
- SaaS Software as a Service: the capability provided to the consumer is to use the provider's applications running on a cloud infrastructure.
- the applications are accessible from various client devices through a thin client interface such as a web browser (e.g., web-based e-mail).
- a web browser e.g., web-based e-mail
- the consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, storage, or even individual application capabilities, with the possible exception of limited user-specific application configuration settings.
- PaaS Platform as a Service
- the consumer does not manage or control the underlying cloud infrastructure including networks, servers, operating systems, or storage, but has control over the deployed applications and possibly application hosting environment configurations.
- IaaS Infrastructure as a Service
- the consumer does not manage or control the underlying cloud infrastructure but has control over operating systems, storage, deployed applications, and possibly limited control of select networking components (e.g., host firewalls).
- Private cloud the cloud infrastructure is operated solely for an organization. It may be managed by the organization or a third party and may exist on-premises or off-premises.
- Public cloud the cloud infrastructure is made available to the general public or a large industry group and is owned by an organization selling cloud services.
- Hybrid cloud the cloud infrastructure is a composition of two or more clouds (private, community, or public) that remain unique entities but are bound together by standardized or proprietary technology that enables data and application portability (e.g., cloud bursting for load-balancing between clouds).
- a cloud computing environment is service oriented with a focus on statelessness, low coupling, modularity, and semantic interoperability.
- An infrastructure that includes a network of interconnected nodes.
- cloud computing environment 50 includes one or more cloud computing nodes 10 with which local computing devices are used by cloud consumers, such as mobile device 54 A, desktop computer 54 B, laptop computer 54 C, and/or automobile computer system 54 N may communicate.
- Nodes 10 may communicate with one another. They may be grouped (not shown) physically or virtually, in one or more networks, such as Private, Community, Public, or Hybrid clouds as described hereinabove, or a combination thereof. This allows cloud computing environment 50 to offer infrastructure, platforms and/or software as services for which a cloud consumer does not need to maintain resources on a local computing device.
- computing devices 54 A-N are intended to be illustrative only and that computing nodes 10 and cloud computing environment 50 can communicate with any type of computerized device over any type of network and/or network addressable connection (e.g., using a web browser).
- FIG. 10 a set of functional abstraction layers provided by cloud computing environment 50 ( FIG. 9 ) is shown. It should be understood in advance that the components, layers, and functions shown in FIG. 10 are intended to be illustrative only and embodiments of the invention are not limited thereto. As depicted, the following layers and corresponding functions are provided:
- Hardware and software layer 60 includes hardware and software components.
- hardware components include: mainframes 61 ; RISC (Reduced Instruction Set Computer) architecture based servers 62 ; servers 63 ; blade servers 64 ; storage devices 65 ; and networks and networking components 66 .
- software components include network application server software 67 and database software 68 .
- Virtualization layer 70 provides an abstraction layer from which the following examples of virtual entities may be provided: virtual servers 71 ; virtual storage 72 ; virtual networks 73 , including virtual private networks; virtual applications and operating systems 74 ; and virtual clients 75 .
- management layer 80 may provide the functions described below.
- Resource provisioning 81 provides dynamic procurement of computing resources and other resources that are utilized to perform tasks within the cloud computing environment.
- Metering and Pricing 82 provide cost tracking as resources are utilized within the cloud computing environment, and billing or invoicing for consumption of these resources. In one example, these resources may include application software licenses.
- Security provides identity verification for cloud consumers and tasks, as well as protection for data and other resources.
- User portal 83 provides access to the cloud computing environment for consumers and system administrators.
- Service level management 84 provides cloud computing resource allocation and management such that required service levels are met.
- Service Level Agreement (SLA) planning and fulfillment 85 provide pre-arrangement for, and procurement of, cloud computing resources for which a future requirement is anticipated in accordance with an SLA.
- SLA Service Level Agreement
- Workloads layer 90 provides examples of functionality for which the cloud computing environment may be utilized. Examples of workloads and functions which may be provided from this layer include: mapping and navigation 91 ; software development and lifecycle management 92 ; virtual classroom education delivery 93 ; data analytics processing 94 ; transaction processing 95 ; and function 96 .
- Function 96 in the present invention is the functionality of adaptive retraining of an artificial intelligence model by detecting data drift, concept drift, and model drift.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A computer-implemented method, a computer program product, and a computer system for adaptive retraining of an artificial intelligence model. A computer system computes drift magnitude scores for respective drift functions. A computer system computes an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift. A computer system computes an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold. A computer system determines whether retraining of the artificial intelligence model is required, based on the overall drift score. A computer system performs the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
Description
- The present invention relates generally to artificial intelligence models, and more particularly to adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift.
- In enterprise artificial intelligence (AI) systems, specifically in industrial Internet of Things (IoT) systems, multiple AI models are prepared and deployed to fulfill business requirements. Monitoring the behavior of these AI models and retrain as the model performance degrades by identifying suitable retraining data is essential to the business. Analysis of the models alone is not sufficient to decide on whether to retrain or to select data for the retraining. It is also required to identify changes in input data (features) and targets (ground truth) to perform retraining. Although various algorithms exist to identify various types of drifts, e.g., a data drift, a concept (target) drift, and a model drift, a system that enables an automated analysis of these drifts and the relationship between them is required for business personnel.
- In one aspect, a computer-implemented method for adaptive retraining of an artificial intelligence model is provided. The computer-implemented method includes computing drift magnitude scores for respective drift functions. The computer-implemented method further includes computing an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift. The computer-implemented method includes computing an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold. The computer-implemented method includes determining whether retraining of the artificial intelligence model is required, based on the overall drift score. The computer-implemented method includes performing the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
- In another aspect, a computer program product for adaptive retraining of an artificial intelligence model is provided. The computer program product comprises a computer readable storage medium having program instructions embodied therewith, and the program instructions are executable by one or more processors. The program instructions are executable to: compute drift magnitude scores for respective drift functions; compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift; compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold; determine whether retraining of the artificial intelligence model is required, based on the overall drift score; and perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
- In yet another aspect, a computer system for adaptive retraining of an artificial intelligence model is provided. The computer system comprises one or more processors, one or more computer readable tangible storage devices, and program instructions stored on at least one of the one or more computer readable tangible storage devices for execution by at least one of the one or more processors. The program instructions are executable to compute drift magnitude scores for respective drift functions. The program instructions are further executable to compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift. The program instructions are further executable to compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold. The program instructions are further executable to determine whether retraining of the artificial intelligence model is required, based on the overall drift score. The program instructions are further executable to perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
-
FIG. 1 is a diagram illustrating a problem setting of a predictive model, in accordance with one embodiment of the present invention. -
FIG. 2 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with one embodiment of the present invention. -
FIG. 3 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. -
FIG. 4 illustrates an example of relationships between a data drift score, a concept drift score, a model drift score, and an overall drift score, in accordance with another embodiment of the present invention. -
FIG. 5 is a diagram illustrating a first example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. -
FIG. 6 is a diagram illustrating a second example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. -
FIG. 7 is a diagram illustrating a third example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. -
FIG. 8 is a diagram illustrating components of a computing device or server, in accordance with one embodiment of the present invention. -
FIG. 9 depicts a cloud computing environment, in accordance with one embodiment of the present invention. -
FIG. 10 depicts abstraction model layers in a cloud computing environment, in accordance with one embodiment of the present invention. - An AI system includes input data, predictive model and parameters, model predictions, ground truth, and various types of drift functions to analyze the data elements and drift thresholds.
FIG. 1 is a diagram illustrating a problem setting ofpredictive model 110, in accordance with one embodiment of the present invention. Input sensor data (X) 120 is generated in periodic interval from multiple sensors installed in various assets and operations. Feature extracted data (X′) 130 is prepared from timeseries of input sensor data (X) 120 using feature extraction techniques. Predictive model or AI model (M) 110 may use either input sensor data (X) 120 or feature extracted data (X′) 130 for training. Predictive model or AI model (M) 110 is trained on input sensor data (X) 120 or feature extracted data (X′) 130 and outputs predictions (Ŷ) 140 in example 1 and/or predictions (Ŷ) 150 in example 2 for all the subsequent input data. In example 1, predictions (Ŷ) 140 is classification output; in example 2, predictions (Ŷ) 150 is regression output. In example 1, ground truth (Y) 160 is available for predictive model (M) 110; in example 2, ground truth (Y) 170 is available for predictive model (M) 110. - The AI system has following drifts. The data drift is a drift in input data. The analysis of the data drift requires only the input data, e.g., timeseries of input sensor data (X) 120 or feature extracted data (X′) 130. The concept drift is a drift in target labels (e.g., ground truth (Y) 160 or ground truth (Y) 170) or the relationship between target labels and input data. The model drift is a drift in predictions of predictive model or AI model (M) 110, the relationship between the predictions (e.g., predictions (Ŷ) 140 and/or predictions (Ŷ) 150) and the target labels, or the relationship between the input data, the predictions, and the target labels. The analysis of the model drift may include comparison of model parameters or optimization parameters emitted during model preparation. A client model (such as predictive model (M) 110) is required in order to analyze the model drift.
- Although the data drift, the concept drift, and the model drift have been explored individually, an automated system, which simultaneously considers all these drift types to perform retraining, does not exist. The data drift has been studied without analyzing model quality. In the data drift driven adaptive retraining, if the drift in input data is identified, a predictive model is retrained anticipating a drift in model performance. The data drift driven adaptive retraining has following drawbacks: (1) The predictive model is retrained without confirming a drift in model predictions. (2) A drift in ground truth (or the concept drift) is not addressed. (3) There is no mechanism to accommodate multiple drift functions from other drift types. Although multiple definitions of the concept drift exist, concept drift driven adaptive retraining, in majority of the cases, is to analyze the target variable. Drawbacks of the concept drift driven adaptive retraining are as follows: (1) The concept drift alone is not the reason for model degradation. (2) There is no mechanism to accommodate multiple drift functions from other drift types. In the quality driven adaptive retraining, model quality is observed over a time period and the abrupt model drift is identified using moving average techniques. In identifying the model quality, model predictions against ground truth are analyzed. The quality driven adaptive retraining has following drawbacks: (1) The reason for the model drift is not explained. (2) There is no mechanism to accommodate multiple drift functions. (3) There is no mechanism or method to anticipate possible model degradation prior to happening. (4) The drift in ground truth (or the concept drift) is not addressed as well.
- Embodiments of the present invention disclose a system and method for quantifying different types of drifts in training artificial intelligence (AI) models, e.g., the data drift, the concept drift, and the model drift, using proper drift functions/algorithms. The disclosed system and method detect the drifts and notify users to retrain the AI models when required, using combined analysis of the above drifts. The disclosed system and method select appropriate data required for retraining using the above analysis and perform retraining of the AI model.
- In embodiments of the present invention, the disclosed system and method train a model on given training data, deploy the model in production, monitor the model over the time, and retrain the model as required. The disclosed system comprises two major modules: a model training module and an adaptive retraining module. The disclosed system and method are implemented on one or more computing devices or servers. A computing device or server is described in more detail in later paragraphs with reference to
FIG. 8 . The disclosed system and method may be implemented in a cloud computing environment. The cloud computing environment is described in more detail in later paragraphs with reference toFIG. 9 andFIG. 10 . - The model training module trains and deploys a predictive model or AI model. From a pool of pre-selected models, the model training module chooses a best model suitable for a given dataset and a problem, using manual analysis or automated selection such as optimization-based selection and grid search. The model training module determines model parameters for the best model, using hyperparameter search. The model training module trains the best model using training data and ground truth. Part of the training data may be used as validation data. The model training module deploys the best model in production. Test data is scored using the deployed best model and predictions are obtained by the model training module. The deployed best model is selected to be monitored by the adaptive retrain module. The implementation of the adaptive retrain module includes two phases:
phase 1—computing retraining score, andphase 2—identifying required retraining data and creating a new AI model. - Embodiments of the present invention disclose an automated system that perform adaptive retraining by analyzing the data drift, the concept drift, and the model drift together. In embodiments of the present invention, an overall drifting score or a retraining score, which considers the relationship between the data drift, the concept drift, and the model drift, is computed. Embodiments of the present invention develop a 3-level drift scoring technique. First, a score for each drift function that resembles magnitude of change occurred in the data is computed. Second, a score for each of the data drift, the concept drift, and the model drift by aggregation of scores from the respective drift functions. Third, the overall drift score (or the retraining score) is computed by considering the second-level scores, i.e., an aggregated data drift score, an aggregated concept drift score, and an aggregated model drift score. Embodiments of the present invention select appropriate retraining data by a combination of sampling and relabeling techniques, based on the 3-level drift scores. Embodiments of the present invention allow users to take an informed decision in automated manner using combined analysis of data, concept, and model drifts.
- In embodiments of the present invention, any algorithm which is used to identify a drift in input data, a drift in ground truth, a drift in model predictions, or relationship among them is called a drift function. Based on the type of data that the algorithm considers, the drift function is categorized into one of the categories: the data drift functions, the concept drift functions, and the model drift functions. The proposed system and method re-categorize all the available drift functions ƒ1, ƒ2, ƒ3, . . . ƒn into data drift functions, concept drift functions, and model drift functions as per drift definitions. Drift thresholds, including a drift threshold (τƒ
i ) for a drift function ƒi, a data drift threshold (τdd), a concept drift threshold (τcd), and a model drift threshold (τmd), can be set either by predefined values or by users. - In computation of the 3-level drift scores, the inputs include a batch of test data, model predictions, and ground truth. The outputs in the first-level drift score computation include a drift magnitude score (sƒ
i ) and a drift flag ζƒi for a drift function ƒi (i=1, . . . , n). The outputs in the second-level drift score computation include an aggregated data drift score (sdd) and its drift flag (ζdd), an aggregated concept drift score (scd) and its drift flag (ζcd), and an aggregated model drift score (ζmd) and its drift flag (ζmd). The outputs in the third-level drift score computation include a retraining score (s) and a retraining flag ζ. The retraining score (s) may also called as an overall drift score or a total drift score. In computation of the 3-level drift scores, the data elements from the model training phase, i.e., training data, validation data, model predictions for training or validation data, ground truth, model parameters, optimization statistics, etc., can be used as a base in the adaptive retrain module. Computation of the 3-level drift scores will be discussed in detail in later paragraphs of this document. - In the third-level drift score computation in phase 1 (computing retrain score) mentioned above, the retraining score (s) and retraining flag ζ are obtained. Based on the retraining score (s) and retraining flag ζ, the disclosed system and method determine whether retraining of predictive model or AI model is required. If the retraining flag ζ is true, the disclosed system and method perform the retraining.
- In the second-level drift score computation in phase 1 (computing retrain score) mentioned above, the aggregated data drift score (sdd), the data drift flag (ζdd), the aggregated concept drift score (scd), the concept drift flag (ζcd), the aggregated model drift score (smd), and the model drift flag (ζcd) are obtained. In response to determining that the data drift flag (ζdd) is true, the disclosed system and method determines that the data drift is present. In response to determining that the concept drift flag (ζcd) is true, the disclosed system and method determines that the concept drift is present. In response to determining that the model drift flag (ζcd) is true, the disclosed system and method determine that the model drift is present.
- In phase 2 (identifying required retraining data and creating a new AI model) mentioned above, in response to determining that the data drift is present, the disclosed system and method select new training data from the drifted period using known sampling techniques. In response to determining that the data drift is not present, the disclosed system and method use the original training data. Different strategies can be used to prepare new training data from the drifted period.
- In phase 2 (identifying required retraining data and creating a new AI model) mentioned above, in response to determining that the concept drift is present, the disclosed system and method apply known relabeling techniques on the selected training data. The disclosed system and method may take new label definitions as inputs or deduce new label definitions from the drifted data.
- In phase 2 (identifying required retraining data and creating a new AI model) mentioned above, in response to determining that the model drift is present and either the concept drift or the data drift is present, model selection is ignored by the disclosed system and method. Otherwise, the disclosed system and method perform model selection to select a new AI model. Finally, hyperparameter search is performed on the selected new AI model.
- The disclosed system and method may replace the previously deployed AI model with the new AI model. Alternatively, the disclosed system and method may compare both the previously deployed AI model with the new AI model against each other using drift results over future period of data, and then the disclosed system and method select a winner as the production model.
- Computation of the first-level drift scores is as follows. A drift function ƒi has a threshold τƒ
i (>0) set to either a user provided value or a default value. ƒi computes metrics for a given batch of data either by comparing against training data or by comparing against recent data over a window or any other means. Example of a metric is standard deviation (std). Finally, a drift magnitude score (sƒi ) and a drift flag (ζƒi ) is computed for ƒi. For example, the drift magnitude score (sƒi ) and drift flag (ζƒi ) are computed by -
- The drift magnitude score (sƒ
i ) and drift flag (ζƒi ) are computed for all data drift functions, concept drift functions, and model drift functions ƒ1, ƒ2, ƒ3, . . . ƒn. - Computation of the second-level drift scores is as follows. Each of the data drift, the concept drift, and the model drift has at least one drift function. The data drift, the concept drift, and the model drift have thresholds τdd, τcd, and τmd, respectively; the thresholds can be set either by predefined values or by users. For example, the data drift has drift function ƒ1, ƒ2, . . . ƒm which evaluate either univariate or multivariate input data. The aggregated data drift score (sdd) and the data drift flag (ζdd) are computed from drift magnitude scores and drift flags corresponding to all the data drift functions ƒ1, ƒ2, . . . ƒm. For example, the aggregated data drift score (sdd) and the data drift flag (ζdd) are computed by
-
- The aggregated concept drift score (scd), the concept drift flag (ζcd), the aggregated model drift score (smd), and model drift flag (ζmd) are computed in a similar way as the aggregated data drift score (sdd) and the data drift flag (ζdd).
- Computation of the third-level drift scores is as follows. The retraining score or overall drift score (s) and the retraining flag ζ are computed based on the aggregated data drift score (sdd) and its drift flag (ζdd), the aggregated concept drift score (scd) and its drift flag (ζcd), and the aggregated model drift score (smd) and its drift flag (ζmd). The aggregated data drift score (sdd), the aggregated concept drift score (scd), and the aggregated model drift score (smd) can be aggregated values over a period. An example of aggregation is the average of the values. For example, the retraining score or overall drift score (s) and the retraining flag ζ are computed by
-
- The above computation is deduced from an approximated relationship between various kinds of drifts.
-
FIG. 2 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with one embodiment of the present invention. The operational steps presented inFIG. 2 are implemented by the disclosed system which is hosted by one or more computing devices or servers. A computing device or server is described in more detail in later paragraphs with reference toFIG. 8 . The operational steps presented inFIG. 2 may be implemented in a cloud computing environment. The cloud computing environment is described in more detail in later paragraphs with reference toFIG. 9 andFIG. 10 . - At
step 201, the one or more computing devices or servers select an artificial intelligence (AI) model for a dataset and a problem. From a pool of pre-selected models, the one or more computing devices or servers choose a best model suitable for the given dataset and the problem. Atstep 202, the one or more computing devices or servers determine parameters for the AI model using hyperparameter search. Atstep 203, the one or more computing devices or servers train the AI model using training data and ground truth. Atstep 204, the one or more computing devices or servers deploy the AI model and obtain predictions using the AI model. Steps 201-204 are operational steps in the model training phase. The data elements from the model training phase, including training data, validation data, model predictions for training or validation data, ground truth, model parameters, and optimization statistics, can be used to compute 3-level drift scores in the following steps. - At
step 205, the one or more computing devices or servers compute drift magnitude scores for respective drift functions. Drift functions ƒ1, ƒ2, ƒ3, . . . ƒn are available, each of which is an algorithm used to identify a drift in input data, a drift in ground truth, a drift in model predictions, or a relationship among them. The one or more computing devices or servers categorize all the available drift functions ƒ1, ƒ2, ƒ3, . . . ƒn into data drift functions, concept drift functions, and model drift functions. For a drift function ƒ1, the one or more computing devices or servers compute a drift magnitude score (sƒi ), usingequation 1 described in a previous paragraph in this document. For a drift function ƒi, the one or more computing devices or servers also compute a drift flag (ζƒi ), usingequation 2 described in a previous paragraph in this document. Atstep 205, the one or more computing devices or servers compute the first-level drift scores. - At
step 206, the one or more computing devices or servers compute an aggregated data drift score, an aggregated concept drift score, and an aggregated model drift score. Based on the drift magnitude score (sƒi ) and the drift flag (ζƒi ) (the first-level drift scores) computed atstep 205, the one or more computing devices or servers compute the aggregated data drift score (sdd), usingequation 3 described in a previous paragraph in this document. The one or more computing devices or servers atstep 206 also compute the data drift flag (ζdd), using equation 4 described in a previous paragraph in this document, based on the aggregated data drift score (sdd) and a predetermined threshold (τdd) of the data shift. Similarly, atstep 206, the one or more computing devices or servers compute the aggregated concept drift score (scd), the concept drift flag (ζcd), the aggregated model drift score (smd), and model drift flag (ζcd). Atstep 206, the one or more computing devices or servers compute the second-level drift scores. - At
step 207, the one or more computing devices or servers compute a retraining score (or an overall drift score), based on the aggregated data drift score, the aggregated concept drift score, and the aggregated model drift score. Based on the aggregated data drift score (sdd), the aggregated concept drift score (scd), and aggregated model drift score (smd) that are computed atstep 206, and also based on the predetermined threshold (τdd) of the data shift, the predetermined threshold (τcd) of the concept shift, and the predetermined threshold (rind) of the model drift, the one or more computing devices or servers compute the retraining score or overall drift score (s), using equations 5-9 described in a previous paragraph in this document. Atstep 207, the one or more computing devices or servers also compute a retraining flag (ζ), usingequation 10 described in a previous paragraph in this document. Atstep 207, the one or more computing devices or servers compute the third-level drift scores. - At
step 208, the one or more computing devices or servers determine whether retaining of the AI model is required, based on the retraining score (the retraining score may also called as an overall drift score or a total drift score). To determine whether the retaining of the AI model is required, the one or more computing devices or servers determines whether the retraining flag ζ is true or false. The retraining score or overall drift score (s) and retraining flag ζ have been computed atstep 207 by the one or more computing devices or servers. - In response to determining that the retaining the AI model is not required (NO branch of decision block 209), the one or more computing devices or servers does not perform the retraining of the AI model. In other words, in response to determining that the retraining flag ζ is false, the one or more computing devices or servers will not take any action to retrain the AI model.
- In response to determining that retaining the AI model is required (YES branch of decision block 209), at
step 210, the one or more computing devices or servers perform the retraining of the AI model. In other words, in response to determining that the retraining flag ζ is true, the one or more computing devices or servers perform the retraining. -
FIG. 3 is a flowchart showing operational steps of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. The operational steps presented inFIG. 3 are implemented by the disclosed system which is hosted by one or more computing devices or servers. A computing device or server is described in more detail in later paragraphs with reference toFIG. 8 . The operational steps may be implemented in a cloud computing environment. The cloud computing environment is described in more detail in later paragraphs with reference toFIG. 9 andFIG. 10 . - The aggregated data drift score, aggregated concept drift score, and aggregated model drift score have been computed at
step 206 by the one or more computing devices or servers. The aggregated data drift score, aggregated concept drift score, and aggregated model drift score are used for the operating steps presented inFIG. 3 . - At
step 301, the one or more computing devices or servers determine whether the data drift is present. To determine whether the data drift is present, the one or more computing devices or servers determine whether the data drift flag (ζdd) (which is computed at step 206) is true or false. If the data drift flag (ζdd) is true, the data drift is present; if the data drift flag (ζdd) is false, the data drift is not present. - In response to determining that the data drift is present or the data drift flag (ζdd) is true (YES branch of decision block 301), at
step 302, the one or more computing devices or servers use new training data for the retraining of the AI model. The new training data may be an entirely new training dataset. The new training data may be combination of the drifted training data and non-drifted training data (or the one or more computing devices or servers incorporate the drifted training data into new data). Selecting the new training data from the drifted period uses known sampling techniques. - In response to determining that the data drift is not present or the drift flag (ζdd) is false (NO branch of decision block 301), at
step 303, the one or more computing devices or servers use original data which has been used in the model training phase. - After either step 302 or step 303, the one or more computing devices or servers at
step 304 determine whether the concept drift is present. To determine whether the concept drift is present, the one or more computing devices or servers determine whether the concept drift flag (ζcd) (which is computed at step 206) is true or false. If the concept drift flag (ζdd) is true, the concept drift is present; if the concept drift flag (ζcd) is false, the concept drift is not present. - In response to determining that the concept drift is present or the concept drift flag (ζcd) is true (YES branch of decision block 304), at
step 305, the one or more computing devices or servers incorporate drifted targets into a dataset for the retraining of the AI model. - In response to determining that the concept drift is not present or the concept drift flag (ζcd) is false (NO branch of decision block 304), at
step 306, the one or more computing devices or servers determine whether the model drift is present. To determine whether the model drift is present, the one or more computing devices or servers determine whether the model drift flag (ζcd) (which is computed at step 206) is true or false. If the model drift flag (ζcd) is true, the model drift is present; if the model drift flag (ζcd) is false, the model drift is not present. - In response to determining that the concept drift is present or the model drift flag (ζcd) is true (YES branch of decision block 306), at
step 307, the one or more computing devices or servers determine whether either data drift or the concept drift is present. In response to determining that the concept drift is not present or the model drift flag (ζcd) is false (NO branch of decision block 306), atstep 308, the one or more computing devices or servers determine whether the concept drift is present. - In response to determining that one of the data drift and the concept drift is present (YES branch of decision block 307), the one or more computing devices or servers at
step 310 perform hyperparameter search. In response to determining that neither the data drift nor the concept drift is present (NO branch of decision block 307), the one or more computing devices or servers atstep 309 perform model selection and then perform hyperparameter search. - In response to determining that the concept drift is present (YES branch of decision block 308), the one or more computing devices or servers at
step 309 perform model selection and then perform hyperparameter search. In response to determining that the concept drift is not present (NO branch of decision block 308), the one or more computing devices or servers will not take any action to retrain the AI model. -
FIG. 4 illustrates an example of relationships between a data drift score, a concept drift score, a model drift score, and an overall drift score, in accordance with another embodiment of the present invention. As shown in the first row inFIG. 4 , when a data drift, a concept drift, and a model drift are present (i.e., data drift flag (ζdd), concept drift flag (ζcd), and model flag (ζmd) are true) and when the overall drift score (which is calculated based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score) indicates that retraining is required, the disclosed system and method incorporate drifted input data and drifted targets into new training data, and the disclosed system and method perform hyperparameter search. - As shown in the second row in
FIG. 4 , when a data drift and a concept drift are present (i.e., data drift flag (ζdd) and concept drift flag (ζcd) are true), when a model drift is not present (model drift flag (ζmd) is false), and when the overall drift score indicates that retraining is required, the disclosed system and method incorporate drifted input data and drifted targets into new training data, the disclosed system and method perform new model selection. - As shown in the third row in
FIG. 4 , when a data drift and a model drift are present (i.e., data drift flag (ζdd) and model drift flag (ζmd) are true), when a concept drift is not present (concept drift flag (ζcd) is false), and when the overall drift score indicates that retraining is required, the disclosed system and method incorporate drifted input data into new training data and perform hyperparameter search. - As shown in the fourth row in
FIG. 4 , when a data drift is not present (data drift flag (ζdd) is false), when a concept drift and a model drift are present (concept drift flag (ζcd), and model drift flag (ζmd) are true), and when the overall drift score indicates that retraining is required, the disclosed system and method incorporate drifted targets into new training data and perform hyperparameter search. - As shown in the fifth row in
FIG. 4 , when a data drift and a model drift are not present (i.e., data drift flag (ζdd) and model flag (ζmd) are false), when a concept drift is present (concept drift flag (ζcd) is true), and when the overall drift score indicates that retraining is required, the disclosed system and method perform new model selection and retrain the new model with new training data along with target relabeling. - As shown in the sixth row in
FIG. 4 , when a data drift and a concept drift are not present (i.e., data drift flag (ζdd) and concept flag (ζcd) are false), when a model drift is present (model drift flag (ζmd) is true), and when the overall drift score indicates that retraining is required, it is indicated that the present deployed model may not be robust. Therefore, the disclosed system and method perform new model selection; for example, the disclosed system and method select a different model from pipeline optimization. - As shown in the seventh row in
FIG. 4 , when a data drift is present (i.e., data drift flag (ζdd) is true), when a concept drift and a model drift are not present (concept drift flag (ζcd) and model drift flag (ζmd) are false), and when the overall drift score indicates that retraining is not required, the disclosed system and method issue a warning of the data drift. - As shown in the eighth row in
FIG. 4 , when a data drift, a concept drift, and a model drift are not present (i.e., data drift flag (ζdd), concept drift flag (ζcd), and model flag (ζmd) are false) and when the overall drift score indicates that retraining is not required, it is indicated that the currently deployed model is in a good shape. Therefore, the disclosed system and method do not take any action for adaptive retraining. -
FIG. 5 is a diagram illustrating a first example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. An AI model M0 is trained on a historic training data, by using model selection and hyperparameter search. M0 is deployed and configured for adaptive retraining with following thresholds: data drift threshold τdd=0.7, concept drift threshold τcd=0.8, and model drift threshold τmd=0.9. Upon arrival of the first batch of production data and when ground truth is available, drift algorithms are applied on the model M0 and the data. The drift scores are computed and obtained as follows: aggregated data drift score sdd=0.8, aggregated concept drift score scd=0.5, aggregated model drift score smd=1.2, and overall drift score (or retraining score) s=1. Further, the computation of drift scores gives data drift flag ζdd=True, concept drift flag ζcd=False, model drift flag ζmd=True, and retraining flag ζ=True. Based on the drift scores and drift flags, M0 is set to retrain and new train data is selected using known sampling techniques. A new AI model M1 is created after hyperparameter search. -
FIG. 6 is a diagram illustrating a second example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. An AI model M0 is trained on a historic training data, by using model selection and hyperparameter search. M0 is deployed and configured for adaptive retraining with following thresholds: data drift threshold τdd=0.7, concept drift threshold τcd=0.8, and model drift threshold τmd=0.9. Upon arrival of the first batch of production data and when ground truth is available, drift algorithms are applied on the model M0 and the data. The drift scores are computed and obtained as follows: aggregated data drift score sdd=0.6, aggregated concept drift score scd=0.9, aggregated model drift score smd=1.2, and overall drift score (or retraining score) s=1. Further, the computation of drift scores gives data drift flag ζdd=False, concept drift flag ζcd=True, model drift flag ζmd=True, and retraining flag ζ=True. Based on the drift scores and drift flags, M0 is set to retrain and new train data is selected after relabeling the ground truth as per new definitions. A new AI model M1 is created after hyperparameter search. -
FIG. 7 is a diagram illustrating a third example of adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift, in accordance with another embodiment of the present invention. An AI model M0 is trained on a historic training data, by using model selection and hyperparameter search. M0 is deployed and configured for adaptive retraining with following thresholds: data drift threshold τdd=0.7, concept drift threshold τcd=0.8, and model drift threshold τmd=0.9. Upon arrival of the first batch of production data and when ground truth is available, drift algorithms are applied on the model M0 and the data. The drift scores are computed and obtained as follows: aggregated data drift score sdd=0.6, aggregated concept drift score scd=0.7, aggregated model drift score smd=0.8, and overall drift score (or retraining score) s=0.8. Further, the computation of drift scores gives data drift flag ζdd=False, concept drift flag ζcd=False, model drift flag ζmd=False, and retraining flag ζ=False. Based on the drift scores and drift flags, retraining is not required. -
FIG. 8 is a diagram illustrating components of computing device orserver 800, in accordance with one embodiment of the present invention, in accordance with one embodiment of the present invention. It should be appreciated thatFIG. 8 provides only an illustration of one implementation and does not imply any limitations; different embodiments may be implemented. - Referring to
FIG. 8 , computing device orserver 800 includes processor(s) 820,memory 810, and tangible storage device(s) 830. InFIG. 8 , communications among the above-mentioned components of computing device orserver 800 are denoted bynumeral 890.Memory 810 includes ROM(s) (Read Only Memory) 811, RAM(s) (Random Access Memory) 813, and cache(s) 815. One or more operating systems 831 and one ormore computer programs 833 reside on one or more computer readable tangible storage device(s) 830. - Computing device or
server 800 further includes I/O interface(s) 850. I/O interface(s) 850 allows for input and output of data with external device(s) 860 that may be connected to computing device orserver 800. Computing device orserver 800 further includes network interface(s) 840 for communications between computing device orserver 800 and a computer network. - The present invention may be a system, a method, and/or a computer program product at any possible technical detail level of integration. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.
- The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.
- Computer readable program instructions described herein can be downloaded to respective computing/processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and/or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and/or edge servers. A network adapter card or network interface in each computing/processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing/processing device.
- Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, and procedural programming languages, such as the C programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.
- Aspects of the present invention are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer readable program instructions.
- These computer readable program instructions may be provided to a processor of a computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and/or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function/act specified in the flowchart and/or block diagram block or blocks.
- The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions/acts specified in the flowchart and/or block diagram block or blocks.
- The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the Figures. For example, two blocks shown in succession may, in fact, be accomplished as one step, executed concurrently, substantially concurrently, in a partially or wholly temporally overlapping manner, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.
- It is to be understood that although this disclosure includes a detailed description on cloud computing, implementation of the teachings recited herein are not limited to a cloud computing environment. Rather, embodiments of the present invention are capable of being implemented in conjunction with any other type of computing environment now known or later developed.
- Cloud computing is a model of service delivery for enabling convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, network bandwidth, servers, processing, memory, storage, applications, virtual machines, and services) that can be rapidly provisioned and released with minimal management effort or interaction with a provider of the service. This cloud model may include at least five characteristics, at least three service models, and at least four deployment models.
- Characteristics are as follows:
- On-demand self-service: a cloud consumer can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with the service's provider.
- Broad network access: capabilities are available over a network and accessed through standard mechanisms that promote use by heterogeneous thin or thick client platforms (e.g., mobile phones, laptops, and PDAs).
- Resource pooling: the provider's computing resources are pooled to serve multiple consumers using a multi-tenant model, with different physical and virtual resources dynamically assigned and reassigned according to demand. There is a sense of location independence in that the consumer generally has no control or knowledge over the exact location of the provided resources but may be able to specify location at a higher level of abstraction (e.g., country, state, or datacenter).
- Rapid elasticity: capabilities can be rapidly and elastically provisioned, in some cases automatically, to quickly scale out and rapidly released to quickly scale in. To the consumer, the capabilities available for provisioning often appear to be unlimited and can be purchased in any quantity at any time.
- Measured service: cloud systems automatically control and optimize resource use by leveraging a metering capability at some level of abstraction appropriate to the type of service (e.g., storage, processing, bandwidth, and active user accounts). Resource usage can be monitored, controlled, and reported, providing transparency for both the provider and consumer of the utilized service.
- Service Models are as follows:
- Software as a Service (SaaS): the capability provided to the consumer is to use the provider's applications running on a cloud infrastructure. The applications are accessible from various client devices through a thin client interface such as a web browser (e.g., web-based e-mail). The consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, storage, or even individual application capabilities, with the possible exception of limited user-specific application configuration settings.
- Platform as a Service (PaaS): the capability provided to the consumer is to deploy onto the cloud infrastructure consumer-created or acquired applications created using programming languages and tools supported by the provider. The consumer does not manage or control the underlying cloud infrastructure including networks, servers, operating systems, or storage, but has control over the deployed applications and possibly application hosting environment configurations.
- Infrastructure as a Service (IaaS): the capability provided to the consumer is to provision processing, storage, networks, and other fundamental computing resources where the consumer is able to deploy and run arbitrary software, which can include operating systems and applications. The consumer does not manage or control the underlying cloud infrastructure but has control over operating systems, storage, deployed applications, and possibly limited control of select networking components (e.g., host firewalls).
- Deployment Models are as follows:
- Private cloud: the cloud infrastructure is operated solely for an organization. It may be managed by the organization or a third party and may exist on-premises or off-premises.
- Community cloud: the cloud infrastructure is shared by several organizations and supports a specific community that has shared concerns (e.g., mission, security requirements, policy, and compliance considerations). It may be managed by the organizations or a third party and may exist on-premises or off-premises.
- Public cloud: the cloud infrastructure is made available to the general public or a large industry group and is owned by an organization selling cloud services.
- Hybrid cloud: the cloud infrastructure is a composition of two or more clouds (private, community, or public) that remain unique entities but are bound together by standardized or proprietary technology that enables data and application portability (e.g., cloud bursting for load-balancing between clouds).
- A cloud computing environment is service oriented with a focus on statelessness, low coupling, modularity, and semantic interoperability. At the heart of cloud computing is an infrastructure that includes a network of interconnected nodes.
- Referring now to
FIG. 9 , illustrativecloud computing environment 50 is depicted. As shown,cloud computing environment 50 includes one or morecloud computing nodes 10 with which local computing devices are used by cloud consumers, such asmobile device 54A,desktop computer 54B,laptop computer 54C, and/orautomobile computer system 54N may communicate.Nodes 10 may communicate with one another. They may be grouped (not shown) physically or virtually, in one or more networks, such as Private, Community, Public, or Hybrid clouds as described hereinabove, or a combination thereof. This allowscloud computing environment 50 to offer infrastructure, platforms and/or software as services for which a cloud consumer does not need to maintain resources on a local computing device. It is understood that the types ofcomputing devices 54A-N are intended to be illustrative only and thatcomputing nodes 10 andcloud computing environment 50 can communicate with any type of computerized device over any type of network and/or network addressable connection (e.g., using a web browser). - Referring now to
FIG. 10 , a set of functional abstraction layers provided by cloud computing environment 50 (FIG. 9 ) is shown. It should be understood in advance that the components, layers, and functions shown inFIG. 10 are intended to be illustrative only and embodiments of the invention are not limited thereto. As depicted, the following layers and corresponding functions are provided: - Hardware and
software layer 60 includes hardware and software components. Examples of hardware components include:mainframes 61; RISC (Reduced Instruction Set Computer) architecture basedservers 62;servers 63;blade servers 64;storage devices 65; and networks andnetworking components 66. In some embodiments, software components include networkapplication server software 67 anddatabase software 68. -
Virtualization layer 70 provides an abstraction layer from which the following examples of virtual entities may be provided:virtual servers 71;virtual storage 72;virtual networks 73, including virtual private networks; virtual applications andoperating systems 74; andvirtual clients 75. - In one example,
management layer 80 may provide the functions described below.Resource provisioning 81 provides dynamic procurement of computing resources and other resources that are utilized to perform tasks within the cloud computing environment. Metering andPricing 82 provide cost tracking as resources are utilized within the cloud computing environment, and billing or invoicing for consumption of these resources. In one example, these resources may include application software licenses. Security provides identity verification for cloud consumers and tasks, as well as protection for data and other resources.User portal 83 provides access to the cloud computing environment for consumers and system administrators.Service level management 84 provides cloud computing resource allocation and management such that required service levels are met. Service Level Agreement (SLA) planning andfulfillment 85 provide pre-arrangement for, and procurement of, cloud computing resources for which a future requirement is anticipated in accordance with an SLA. -
Workloads layer 90 provides examples of functionality for which the cloud computing environment may be utilized. Examples of workloads and functions which may be provided from this layer include: mapping andnavigation 91; software development andlifecycle management 92; virtualclassroom education delivery 93; data analytics processing 94;transaction processing 95; andfunction 96.Function 96 in the present invention is the functionality of adaptive retraining of an artificial intelligence model by detecting data drift, concept drift, and model drift.
Claims (20)
1. A computer-implemented method for adaptive retraining of an artificial intelligence model, the method comprising:
computing drift magnitude scores for respective drift functions;
computing an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift;
computing an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold;
determining whether retraining of the artificial intelligence model is required, based on the overall drift score; and
performing the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
2. The computer-implemented method of claim 1 , further comprising:
computing a data drift flag based on the aggregated data drift score and the predetermined data drift threshold, a concept drift flag based on the aggregated concept drift score and the predetermined concept drift threshold, and a model drift flag based on the aggregated model drift score and the predetermined model drift threshold.
3. The computer-implemented method of claim 1 , further comprising:
determining whether the data drift is present, by determining whether a data drift flag is true or false;
in response to determining that the data drift is present, using new training data to retrain the artificial intelligence model; and
in response to determining that the data drift is not present, using original training data.
4. The computer-implemented method of claim 1 , further comprising:
determining whether the concept drift is present, by determining whether a concept drift flag is true or false; and
in response to determining that the concept drift is present, incorporating drifted targets into a dataset for the retraining of the artificial intelligence model.
5. The computer-implemented method of claim 1 , further comprising:
determining whether the model drift is present, by determining whether a model drift flag is true or false;
in response to determining that the model drift is present, determining whether either the data drift or the concept drift is present; and
in response to determining that the model drift is not present, determining whether the concept drift is present.
6. The computer-implemented method of claim 5 , further comprising:
in response to determining that the model drift is present and determining that either the data drift or the concept drift is present, performing hyperparameter search in the retraining; and
in response to determining that the model drift is present and determining that neither the data drift nor the concept drift is present, performing model selection to select a new artificial intelligence model and performing hyperparameter search in the retraining.
7. The computer-implemented method of claim 5 , further comprising:
in response to determining that the model drift is not present and determining that the concept drift is present, performing model selection to select a new artificial intelligence model and performing hyperparameter search in the retraining.
8. A computer program product for adaptive retraining of an artificial intelligence model, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by one or more processors, the program instructions executable to:
compute drift magnitude scores for respective drift functions;
compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift;
compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold;
determine whether retraining of the artificial intelligence model is required, based on the overall drift score; and
perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
9. The computer program product of claim 8 , further comprising the program instructions executable to:
compute a data drift flag based on the aggregated data drift score and the predetermined data drift threshold, a concept drift flag based on the aggregated concept drift score and the predetermined concept drift threshold, and a model drift flag based on the aggregated model drift score and the predetermined model drift threshold.
10. The computer program product of claim 8 , further comprising the program instructions executable to:
determine whether the data drift is present, by determining whether a data drift flag is true or false;
in response to determining that the data drift is present, use new training data to retrain the artificial intelligence model; and
in response to determining that the data drift is not present, use original training data.
11. The computer program product of claim 8 , further comprising the program instructions executable to:
determine whether the concept drift is present, by determining whether a concept drift flag is true or false; and
in response to determining that the concept drift is present, incorporate drifted targets into a dataset for the retraining of the artificial intelligence model.
12. The computer program product of claim 8 , further comprising the program instructions executable to:
determine whether the model drift is present, by determining whether a model drift flag is true or false;
in response to determining that the model drift is present, determine whether either the data drift or the concept drift is present; and
in response to determining that the model drift is not present, determine whether the concept drift is present.
13. The computer program product of claim 12 , further comprising program instructions executable to:
in response to determining that the model drift is present and determining that either the data drift or the concept drift is present, perform hyperparameter search in the retraining; and
in response to determining that the model drift is present and determining that neither the data drift nor the concept drift is present, perform model selection to select a new artificial intelligence model and perform hyperparameter search in the retraining.
14. The computer program product of claim 12 , further comprising the program instructions executable to:
in response to determining that the model drift is not present and determining that the concept drift is present, perform model selection to select a new artificial intelligence model and perform hyperparameter search in the retraining.
15. A computer system for adaptive retraining of an artificial intelligence model, the computer system comprising one or more processors, one or more computer readable tangible storage devices, and program instructions stored on at least one of the one or more computer readable tangible storage devices for execution by at least one of the one or more processors, the program instructions executable to:
compute drift magnitude scores for respective drift functions;
compute an aggregated data drift score for a data drift, an aggregated concept drift score for a concept drift, and an aggregated model drift score for a model drift;
compute an overall drift score, based on the aggregated data drift score, the aggregated concept drift score, the aggregated model drift score, a predetermined data drift threshold, a predetermined concept drift threshold, and a predetermined model drift threshold;
determine whether retraining of the artificial intelligence model is required, based on the overall drift score; and
perform the retraining of the artificial intelligence model, in response to determining the retraining of the artificial intelligence model is required.
16. The computer system of claim 15 , further comprising the program instructions executable to:
compute a data drift flag based on the aggregated data drift score and the predetermined data drift threshold, a concept drift flag based on the aggregated concept drift score and the predetermined concept drift threshold, and a model drift flag based on the aggregated model drift score and the predetermined model drift threshold.
17. The computer system of claim 15 , further comprising the program instructions executable to:
determine whether the data drift is present, by determining whether a data drift flag is true or false;
in response to determining that the data drift is present, use new training data to retrain the artificial intelligence model; and
in response to determining that the data drift is not present, use original training data.
18. The computer system of claim 15 , further comprising the program instructions executable to:
determine whether the concept drift is present, by determining whether a concept drift flag is true or false; and
in response to determining that the concept drift is present, incorporate drifted targets into a dataset for the retraining of the artificial intelligence model.
19. The computer system of claim 15 , wherein further comprising the program instructions executable to:
determine whether the model drift is present, by determining whether a model drift flag is true or false;
in response to determining that the model drift is present, determine whether either the data drift or the concept drift is present; and
in response to determining that the model drift is not present, determine whether the concept drift is present.
20. The computer system of claim 19 , further comprising program instructions executable to:
in response to determining that the model drift is present and determining that either the data drift or the concept drift is present, perform hyperparameter search in the retraining;
in response to determining that the model drift is present and determining that neither the data drift nor the concept drift is present, perform model selection to select a new artificial intelligence model and perform hyperparameter search in the retraining; and
in response to determining that the model drift is not present and determining that the concept drift is present, perform model selection to select the new artificial intelligence model and perform hyperparameter search in the retraining.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/663,848 US20230376825A1 (en) | 2022-05-18 | 2022-05-18 | Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/663,848 US20230376825A1 (en) | 2022-05-18 | 2022-05-18 | Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift |
Publications (1)
Publication Number | Publication Date |
---|---|
US20230376825A1 true US20230376825A1 (en) | 2023-11-23 |
Family
ID=88791782
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/663,848 Pending US20230376825A1 (en) | 2022-05-18 | 2022-05-18 | Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift |
Country Status (1)
Country | Link |
---|---|
US (1) | US20230376825A1 (en) |
-
2022
- 2022-05-18 US US17/663,848 patent/US20230376825A1/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11062230B2 (en) | Detecting data anomalies | |
US11301315B2 (en) | Automated hardware failure prediction framework | |
US11507716B2 (en) | Predicting life expectancy of machine part | |
US20220188690A1 (en) | Machine learning security threat detection using a meta-learning model | |
WO2021171126A1 (en) | Personalized automated machine learning | |
US20230325397A1 (en) | Artificial intelligence based problem descriptions | |
US20210110248A1 (en) | Identifying and optimizing skill scarcity machine learning algorithms | |
US20200272973A1 (en) | Root Cause Identification and Analysis | |
US11915106B2 (en) | Machine learning for determining suitability of application migration from local to remote providers | |
US11449772B2 (en) | Predicting operational status of system | |
US11410049B2 (en) | Cognitive methods and systems for responding to computing system incidents | |
WO2023094916A1 (en) | Identifying persistent anomalies for failure prediction | |
WO2023072443A1 (en) | Workload migration | |
US20200410387A1 (en) | Minimizing Risk Using Machine Learning Techniques | |
US11948101B2 (en) | Identification of non-deterministic models of multiple decision makers | |
US20230117225A1 (en) | Automated workflow analysis and solution implementation | |
US11256598B2 (en) | Automated selection of performance monitors | |
US11212162B2 (en) | Bayesian-based event grouping | |
US10831638B2 (en) | Automated analytics for improving reuse of application solutions | |
US10949764B2 (en) | Automatic model refreshment based on degree of model degradation | |
US20230267323A1 (en) | Generating organizational goal-oriented and process-conformant recommendation models using artificial intelligence techniques | |
US20230274160A1 (en) | Automatically training and implementing artificial intelligence-based anomaly detection models | |
US20230376825A1 (en) | Adaptive retraining of an artificial intelligence model by detecting a data drift, a concept drift, and a model drift | |
US20230137184A1 (en) | Incremental machine learning for a parametric machine learning model | |
US20220013239A1 (en) | Time-window based attention long short-term memory network of deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: INTERNATIONAL BUSINESS MACHINES CORPORATION, NEW YORK Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GANAPAVARAPU, VENKATA SITARAMAGIRIDHARGANESH;YEO, KYONG MIN;ZHOU, NIANJUN;AND OTHERS;REEL/FRAME:059941/0720 Effective date: 20220517 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: DOCKETED NEW CASE - READY FOR EXAMINATION |